AI model rankings

How to sort this table

The default order is by LLM Index Score, which weights seven capabilities into one figure. That is the right starting point and the wrong finishing point: your workload almost certainly does not weight those capabilities the way the overall score does. Sort by an individual category column when the job is narrow — code generation, long-context retrieval, quantitative work — and the ordering will change, sometimes substantially.

Sorting by price is the other view worth taking. Capability and cost are set independently, and the difference between the first and fifth model is frequently a rounding error next to the difference in what they charge. For sustained or high-volume work, the model a few places down the table is often the correct answer.

Why some rows have no score

Of 338 indexed models, 26 carry a computed score. The rest are absent from the current benchmark snapshot — usually because they shipped after it was taken. Those rows keep their factual data and are labelled explicitly rather than being silently sorted to the bottom, which would read as "measured and found wanting" when the truth is "not measured yet".

We do not fill those gaps with estimates. Interpolating a score from a model's family, its parameter count or its vendor's track record would produce a number that looks exactly like a measurement and is not one. When the next snapshot lands, those rows gain real scores.

What the columns mean

The seven category columns are the components of the overall score, each a plain mean of its raw LiveBench tasks on a 0–100 scale. Price is per million input tokens from live provider data; output tokens are priced separately and are usually the larger share of a real bill, so check the model page before budgeting. Context is the maximum window accepted in one request, covering prompt and response together. Neither price nor context feeds the score — they sit beside it because a ranking without cost is only half a decision. Full derivation is on the methodology page.